Fable 5.2 benchmarks, itemized: 90% of Anthropic code now Claude-written
anitakirkovska · reddit · 2026-11-10
A benchmark-by-benchmark read of Anthropic's Fable 5.2, its 13th release this year (2.6x more launches than last year). Key takeaways:
- 4/10 freelance jobs now reliably doable, up from <2/10
- Coordinated 1,000+ agents over 8 days to build an OS that boots on a laptop
- Half of sales/marketing/finance/support/ops tasks can be automated end-to-end
- 2x costlier per token than Opus 5.5, but same coding work costs nearly half — optimize for cost per completed task
- CAD score nearly doubled (44.4 → 83.5); hardware prototyping may be the next startup wave
- Legal still unreliable: only 7.2% passed all checks, human review required
- Over 90% of merged changes in Anthropic's core codebase are now Claude-written
- Computer-use at 89% on tested OSWorld tasks
Related event: Anthropic Releases Claude Fable 5.2 and Mythos 5.2(77 posts)→
More from AGI Musings
- 19-year-old on the futility of convincing people AI isn't the enemy — TechRenamed · 2026-11-10
- Understanding is modeling, thinking is graph search: a compact AI framing — attentionmech · 2026-11-10
- Jaron Lanier in NYT: agents are 'rituals of belief,' Valley self-sabotage — lrettig · 2026-11-10
- Warren McCulloch, father of neural networks, talks AI in rare 1962 footage — burny_tech · 2026-11-10
- Dev mocks tone-filtering AI: like a phone refusing to send unless it sounds emotionally balanced — PierceLilholt · 2026-11-10
- AI circles debate RSI: measuring recursive self-improvement 'turns it dumb' — repligate · 2026-11-10